FTR, a simple RPi4 yields ~4FPS on a DL object detection model. You may need cloud for your central hub, but neural network inference can be done locally.